A priori reconstruction of Thermal-Large Eddy Simulation (T-LES) by Deep Learning Reconstruction a priori de champs de Simulations des Grandes Echelles Thermiques par Apprentissage Profond

In this paper, we examine a machine learning-based method aimed at improving the accuracy of T-LES fields in the context of highly anisothermal flows. We compare this method with an already existing super-resolution method. We train our convolutional neural network by filtering Direct Numerical Simulation (DNS) snapshots into T-LES ones, and optimize our network to reconstruct DNS small scales from T-LES snapshots. Our results show that the neural network outperforms the classical reconstruction method in terms of the quality of the reconstructed coherent structures, but ends up increasing the Root Mean Square (RMS) values over the DNS ones.

In this paper, we examine a machine learning-based method aimed at improving the accuracy of T-LES fields in the context of highly anisothermal flows. We compare this method with an already existing super-resolution method. We train our convolutional neural network by filtering Direct Numerical Simulation (DNS) snapshots into T-LES ones, and optimize our network to reconstruct DNS small scales from T-LES snapshots. Our results show that the neural network outperforms the classical reconstruction method in terms of the quality of the reconstructed coherent structures, but ends up increasing the Root Mean Square (RMS) values over the DNS ones.

Anisothermal flow Deep Learning Super-resolution Heat transfer Thermal-Large Eddy Simulations

Anisothermal flow Deep Learning Super-resolution Heat transfer Thermal-Large Eddy Simulations